Detection of Bowel Sounds Kaja Kvello Master of Science in Cybernetics and Robotics Submission date: June 2018 Supervisor: Øyvind Stavdahl, ITK Co-supervisor: Konstanze Kölle, ITK Anders Lyngvi Fougner, ITK Norwegian University of Science and Technology Department of Engineering Cybernetics Abstract This thesis describes the development and implementation of a bowel sound detection algorithm, which ultimately might be used as part of a meal detection system in an arti- ficial pancreas. An artificial pancreas is an advanced blood glucose regulation system for patients with diabetes. In order to correctly provide insulin doses to the user, artificial pan- creases need to detect when the patient is eating. One possible approach for meal detection is monitoring bowel sound occurrence frequency. This further necessitates a bowel sound detection system, such as the one described in this thesis. The detection algorithm will consist of a combination of signal processing techniques for feature extraction, and pattern recognition in order to classify sound segments as either bowel sound segments or non-bowel sound segments. A linear support vector machine was chosen for classification in the final solution. The Python implementation of the detection system yielded promising results on a provided data set of sound measurements from the stomach. However, certain improvements and refinements should be done to the system before potential integration into an artificial pancreas. The thesis also briefly discusses these further development needs. Chapter1 explains the motivation behind the assignment, as well as the problem descrip- tion and approach. Chapter2 provides the most important theoretical background for the work described in this thesis. Chapters3 and4 describe the method and work towards the final product. These chapters also contain intermediate results which constitute the basis for decisions made throughout the process. Results obtained from the final solution can be found in chapter5, while chapter6 summarizes and concludes the report. i ii Sammendrag Denne oppgaven beskriver utvikling og implementasjon av en deteksjonsalgoritme for tarmlyder, som muligens kan brukes som en del av et maltiddeteksjonssystem˚ i en kunstig bukspyttkjertel. En kunstig bukspyttkjertel er et avansert blodsukkerreguleringssystem for pasienter med diabetes. For a˚ korrekt tilføre insulindoser til brukeren, ma˚ kunstige buk- spyttkjertler oppdage nar˚ pasienten spiser. En mulig metode for maltidsdeteksjon˚ er a˚ overvake˚ forekomstfrekvensen av tarmlyder. Dette krever videre et deteksjonssystem for tarmlyder, slik som beskrevet i denne avhandlingen. Deteksjonsalgoritmen vil besta˚ av en kombinasjon av signalbehandlingsteknikker for egen- skapsuttrekking, samt mønstergjenkjenning for a˚ klassifisere lydsegmenter som enten tarmlydsegmenter eller ikke-tarmlydsegmenter. En lineær støttervektormaskin (support vector machine) ble valgt for klassifisering i den endelige løsningen. Python-implementasjonen av deteksjonssystemet ga lovende resultater pa˚ et gitt datasett med lydmalinger˚ fra magen. Allikevel er det nødvendig med en viss grad av raffinering og forbedring av systemet før eventuell integrering i en kunstig bukspyttkjertel. Rapporten drøfter ogsa˚ disse videre utviklingsbehovene. Kapittel ?? forklarer motivasjonen bak oppgaven, samt oppgavebeskrivelsen og tilnrmin- gen til problemet. Kapittel2 gir den viktigste teoretiske bakgrunnen for arbeidet som er beskrevet i denne oppgaven. Kapitlene3 og4 beskriver metodene brukt i arbeidet mot sluttproduktet. Disse kapitlene inneholder ogs mellomliggende resultater som dan- ner grunnlag for beslutninger som er gjort gjennom hele prosessen. Resultatene fra den endelige lsningen finner du i kapittel5, mens kapittel6 oppsummerer og konkluderer rap- porten. iii iv Preface This master thesis is submitted as a part of the requirements for the M.Sc. degree within the field of Engineering Cybernetics at the Norwegian University of Science and Technol- ogy (NTNU). The assignment was initiated by the Artificial Pancreas Trondheim (APT) research group, and the thesis was written for the Department of Engineering Cybernetics at NTNU. This thesis is a derivative of the ongoing Ph. D work of Konstanze Kolle,¨ and was carried out during spring 2017. The aforementioned Ph. D thesis is concerned with meal detection in artificial pancreas systems, and revealed the potential benefit of a bowel sound detection system as a part of the total meal detection solution. The APT group provided the data set of sound measurements as a .wav-file, and pro- vided the problem description, which asked for detection of the bowel sounds in the data set. Pattern recognition was mentioned as a possible method, but the final problem ap- proach presented in section 1.3 was my own idea. My co-supervisor, Konstanze Kolle,¨ provided relevant background literature in the initial stages of the work. Throughout the process she has also contributed with thoughts and discussion on my various ideas and suggestions for problem specification, solution methods and report structure. My main supervisor, Øyvind Stavdahl also came with brief input on the form of the report and with initial suggestions regarding signal processing methods. The Department of Engineering Cybernetics provided a work station and computer, and APT conveyed a wish of imple- mentation in Python. Apart from this, no assistance was given in the fulfillment of this thesis. To not exploit the resources in the research group, no experts were hired to label the bowel sounds in the data set. All work of manual data marking was done by me, as explained in section 4.4. I would like to thank Konstanze Kolle¨ for all her guidance throughout this semester. Also Øyvind Stavdahl for his comments and help from across the globe. Further, I want to thank Anja Bøe for sharing her medical knowledge of bowel sounds and Vincent Simensen, both for help with proofreading and for comfort in my times of discouragement. Lastly, thanks to my parents for their unwavering support throughout my student years. v vi TABLE OF CONTENTS Abstracti Sammendrag iii Prefacev Table of Contents vii List of Tables xi List of Figures xiii Abbreviations xv 1 Introduction1 1.1 Motivation..................................1 1.1.1 Diabetes...............................1 1.1.2 Artificial pancreas solutions....................3 1.1.3 Artificial Pancreas Trondheim...................4 1.1.4 Bowel sound detection in an artificial pancreas system......4 1.2 Objective..................................5 1.2.1 Given problem...........................5 1.2.2 Interpretation of the problem text.................6 1.3 Approach..................................6 vii 2 Background9 2.1 Bowel sounds................................9 2.1.1 Physiological description......................9 2.1.2 Acquisition of data.........................9 2.1.3 Bowel sound signals........................ 10 2.2 Signal processing methods......................... 13 2.2.1 Frequency analysis......................... 13 2.2.2 Signal filtering........................... 16 2.2.3 Hilbert transform.......................... 18 2.3 Pattern recognition............................. 19 2.3.1 A generic classification system................... 19 2.3.2 Unsupervised learning methods.................. 21 2.3.3 Supervised learning methods.................... 22 2.3.4 Evaluation of classifiers...................... 25 2.4 Feature selection.............................. 27 2.4.1 Recursive feature elimination................... 27 2.5 Python.................................... 28 2.5.1 SciPy................................ 28 3 Feature Extraction 31 3.1 Validation of bowel sound characteristics................. 31 3.1.1 Frequency spectrum........................ 32 3.1.2 Amplitude............................. 38 3.1.3 Sound duration........................... 39 3.2 Extraction of features............................ 40 3.2.1 Segmentation of data........................ 40 3.2.2 Choice of features and signal processing methods......... 42 3.2.3 Calculation of features....................... 43 4 Recognizing Bowel Sounds 47 4.1 An overview of the classification system.................. 47 4.2 Construction and splitting of data set.................... 48 4.3 Choice of pattern recognition method................... 48 4.3.1 Test of unsupervised algorithm................... 48 4.3.2 Supervised algorithm........................ 52 4.4 Marking the data set............................. 52 4.5 Feature selection.............................. 53 5 Results 57 5.1 Classification without feature selection................... 57 5.1.1 Classifier metrics.......................... 57 5.1.2 Visual evaluation of validation set................. 59 5.2 Classification with feature selection.................... 63 5.2.1 Classifier metrics.......................... 63 5.2.2 Visual evaluation of validation set................. 64 viii 6 Conclusion 67 6.1 Further work................................ 68 Bibliography 71 ix x LIST OF TABLES 2.1 Bowel sound (BS) characteristics found in literature............ 12 2.2 Confusion matrix for a binary classification problem........... 26 3.1 Table of features............................... 43 4.1 Features chosen by RFECV in the initial execution............. 54 4.2 Final features chosen by RFECV in the second execution.......... 55 5.1 Confusion matrix calculated from evaluation of the classifier without fea- ture selection................................ 58 5.2 Evaluation metrics for the SVM classifier without